zai-org/glm-4-9b

1import regex as re2import base643import os4import tiktoken5from typing import List, Optional, Union, Dict6from transformers import PreTrainedTokenizer7from transformers.utils import PaddingStrategy8from transformers.tokenization_utils_base import EncodedInput, BatchEncoding9 10 11class ChatGLM4Tokenizer(PreTrainedTokenizer):12    vocab_files_names = {"vocab_file": "tokenizer.model"}13    model_input_names = ["input_ids", "attention_mask", "position_ids"]14 15    def __init__(16            self,17            vocab_file,18            clean_up_tokenization_spaces=False,19            **kwargs20    ):21        self.name = "GLM4Tokenizer"22        self.vocab_file = vocab_file23        pat_str = "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"24        self.pat_str = re.compile(pat_str)25 26        mergeable_ranks = {}27        with open(vocab_file) as f:28            for line in f:29                token, rank = line.strip().split()30                rank = int(rank)31                token = base64.b64decode(token)32                mergeable_ranks[token] = rank33 34        self.mergeable_ranks = mergeable_ranks35 36        self.tokenizer = tiktoken.Encoding(37            name="my_tokenizer",38            pat_str=pat_str,39            mergeable_ranks=mergeable_ranks,40            special_tokens={}41        )42        self.decoder = {rank: token for token, rank in mergeable_ranks.items()}43        self.n_words = len(self.decoder)44 45        super().__init__(46            clean_up_tokenization_spaces=clean_up_tokenization_spaces,47            **kwargs48        )49 50    @property51    def vocab_size(self):52        return self.n_words53 54    def get_vocab(self):55        """ Returns vocab as a dict """56        vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)}57        vocab.update(self.added_tokens_encoder)58        return vocab59 60    def convert_tokens_to_string(self, tokens: List[Union[bytes, str, int]]) -> str:61        """62        Converts a sequence of tokens in a single string.63        """64        text = ""65        temp = b""66        for t in tokens:67            if isinstance(t, int):68                t = chr(t)69            if isinstance(t, str):70                if temp:71                    text += temp.decode("utf-8", errors="replace")72            elif isinstance(t, bytes):73                temp += t74            else:75                raise TypeError("token should only be of type int, bytes or str")76        if temp:77            text += temp.decode("utf-8", errors="replace")78        return text79 80    def _tokenize(self, text, **kwargs):81        tokens = []82        ids = self.tokenizer.encode(text)83        for t in ids:84            tokens.append(self.decoder[t])85        return tokens86 87    def _convert_token_to_id(self, token):88        """ Converts a token (str) in an id using the vocab. """89        return self.mergeable_ranks[token]90 91    def _convert_id_to_token(self, index):92        """Converts an index (integer) in a token (str) using the vocab."""93        return self.decoder.get(index, "")94 95    def save_vocabulary(self, save_directory, filename_prefix=None):96        """97        Save the vocabulary and special tokens file to a directory.98 99        Args:100            save_directory (`str`):101                The directory in which to save the vocabulary.102            filename_prefix (`str`, *optional*):103                An optional prefix to add to the named of the saved files.104 105        Returns:106            `Tuple(str)`: Paths to the files saved.107        """108        if os.path.isdir(save_directory):109            vocab_file = os.path.join(110                save_directory, self.vocab_files_names["vocab_file"]111            )112        else:113            vocab_file = save_directory114 115        with open(self.vocab_file, 'rb') as fin:116            proto_str = fin.read()117 118        with open(vocab_file, "wb") as writer:119            writer.write(proto_str)120 121        return (vocab_file,)122 123    def get_prefix_tokens(self):124        prefix_tokens = [self.convert_tokens_to_ids("[gMASK]"), self.convert_tokens_to_ids("<sop>")]125        return prefix_tokens126 127    def build_single_message(self, role, metadata, message, tokenize=True):128        assert role in ["system", "user", "assistant", "observation"], role129        if tokenize:130            role_tokens = [self.convert_tokens_to_ids(f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n",131                                                                                              disallowed_special=())132            message_tokens = self.tokenizer.encode(message, disallowed_special=())133            tokens = role_tokens + message_tokens134            return tokens135        else:136            return str(f"<|{role}|>{metadata}\n{message}")137 138    def build_inputs_with_special_tokens(139            self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None140    ) -> List[int]:141        """142        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and143        adding special tokens. A BERT sequence has the following format:144 145        - single sequence: `[CLS] X [SEP]`146        - pair of sequences: `[CLS] A [SEP] B [SEP]`147 148        Args:149            token_ids_0 (`List[int]`):150                List of IDs to which the special tokens will be added.151            token_ids_1 (`List[int]`, *optional*):152                Optional second list of IDs for sequence pairs.153 154        Returns:155            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.156        """157        prefix_tokens = self.get_prefix_tokens()158        token_ids_0 = prefix_tokens + token_ids_0159        if token_ids_1 is not None:160            token_ids_0 = token_ids_0 + token_ids_1 + [self.convert_tokens_to_ids("<eos>")]161        return token_ids_0162 163    def _pad(164            self,165            encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],166            max_length: Optional[int] = None,167            padding_side: str = "left",168            padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,169            pad_to_multiple_of: Optional[int] = None,170            return_attention_mask: Optional[bool] = None,171    ) -> dict:172        """173        Pad encoded inputs (on left/right and up to predefined length or max length in the batch)174 175        Args:176            encoded_inputs:177                Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).178            max_length: maximum length of the returned list and optionally padding length (see below).179                Will truncate by taking into account the special tokens.180            padding_strategy: PaddingStrategy to use for padding.181 182                - PaddingStrategy.LONGEST Pad to the longest sequence in the batch183                - PaddingStrategy.MAX_LENGTH: Pad to the max length (default)184                - PaddingStrategy.DO_NOT_PAD: Do not pad185                The tokenizer padding sides are defined in self.padding_side:186 187                    - 'left': pads on the left of the sequences188                    - 'right': pads on the right of the sequences189            pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.190                This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability191                `>= 7.5` (Volta).192            return_attention_mask:193                (optional) Set to False to avoid returning attention mask (default: set to model specifics)194        """195        # Load from model defaults196 197        required_input = encoded_inputs[self.model_input_names[0]]198        seq_length = len(required_input)199 200        if padding_strategy == PaddingStrategy.LONGEST:201            max_length = len(required_input)202 203        if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):204            max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of205 206        needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length207 208        # Initialize attention mask if not present.209        if "attention_mask" not in encoded_inputs:210            encoded_inputs["attention_mask"] = [1] * seq_length211 212        if "position_ids" not in encoded_inputs:213            encoded_inputs["position_ids"] = list(range(seq_length))214 215        if needs_to_be_padded:216            difference = max_length - len(required_input)217 218            if "attention_mask" in encoded_inputs:219                encoded_inputs["attention_mask"] = [0] * difference + encoded_inputs["attention_mask"]220            if "position_ids" in encoded_inputs:221                encoded_inputs["position_ids"] = [0] * difference + encoded_inputs["position_ids"]222            encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input223 224        return encoded_inputs